Improved YOLO for long range detection of small drones.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Accurate drone detection is vital for public safety but challenged by complex environments and limited target features.
• Deep learning models often require significant computational resources, hindering deployment on low-capacity devices.
• Existing methods struggle with small, distant drone detection and efficiency.

Purpose of the Study:

• To propose LMWP-YOLO, a lightweight and efficient drone detection method.
• To improve feature representation and bounding box accuracy for small targets.
• To enhance computational efficiency for deployment on resource-constrained platforms.

Main Methods:

• Developed LMWP-YOLO, incorporating a multidimensional collaborative attention mechanism and multi-scale fusion.
• Utilized depthwise separable convolutions and efficient activation functions for parameter reduction.
• Implemented an optimized loss function for small target bounding box matching and a pruning strategy for filter optimization.

Main Results:

• LMWP-YOLO achieved a 22.07% increase in mean Average Precision (mAP) compared to YOLO11n.
• The model demonstrated a 52.51% reduction in parameters, enhancing efficiency.
• Showcased strong cross-dataset generalization, balancing detection accuracy and computational performance.

Conclusions:

• LMWP-YOLO offers a significant advancement in lightweight drone detection.
• The method effectively addresses challenges of complex environments and small target identification.
• Provides a practical solution for real-time drone monitoring on low-capacity platforms.